🤖 machine learning

Sample Efficient Generative Optimization for Molecular Design

The paper introduces Sample Efficient Generative Optimization (SEGO), a framework that integrates Bayesian optimization with adaptive generative models to significantly reduce the number of costly oracle evaluations required for molecular design, achieving state-of-the-art performance on practical benchmarks with up to ten times fewer samples than existing methods.

Original authors: Sarina Kopf, Cristina Nevado, Philippe Schwaller

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sarina Kopf, Cristina Nevado, Philippe Schwaller

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a treasure hunter in a vast, foggy archipelago called "Chemical Space." Your goal? To find a specific, magical island (a molecule) that cures a disease or powers a new engine. The problem? The islands are infinite, and checking if an island is the right one requires sending a expensive, slow-moving ship (an experiment or high-fidelity simulation) to visit it. You only have fuel for a handful of trips.

Most treasure hunters use one of two strategies, and both have flaws. Some use Generative Models, which are like magical compasses that spin out random islands. They are good at exploring, but they often spin out useless rocks because they don't learn quickly from their mistakes. Others use Bayesian Optimization, which is like a super-smart mapmaker. The mapmaker draws a map based on the few islands you've visited and guesses where the treasure might be. But this mapmaker is picky: it needs a pre-drawn list of islands to work with, or it gets confused if the islands look too different from the ones it knows.

Enter SEGO (Sample Efficient Generative Optimization), a new hybrid strategy that acts like a smart scout team.

The Scout Team Strategy

SEGO combines the best of both worlds into a loop that learns incredibly fast. Here is how it works, step-by-step:

The Mapmaker (The Surrogate): First, a probabilistic mapmaker looks at the few islands you've already visited. It doesn't just guess; it forms a "hypothesis" about where the treasure is hiding in the fog.
The Scout (The Generative Model): Instead of waiting for a pre-made list, SEGO sends a generative model (a "scout") to that specific foggy region. The scout is steered by the mapmaker's hypothesis to generate a fresh batch of candidate islands right where the treasure is likely to be.
The Selection: From this new batch, a selection tool picks the single most promising island to visit.
The Feedback Loop: You send your ship to that island. The result (the "oracle call") is used for two things:
- It sharpens the mapmaker's map, making the next guess even better.
- It "anchors" the scout, teaching it to stick to the real, high-reward areas rather than wandering off.

Why It's a Game-Changer

The paper shows that this method is incredibly sample efficient, meaning it finds the treasure with far fewer ship trips than anyone else.

The PMO Benchmark: On a standard test called the Practical Molecular Optimization (PMO) benchmark, SEGO found the best candidates using only one-tenth of the ship trips (oracle calls) that other methods needed. While others might need 1,000 trips, SEGO found top-tier results with just 100.
The Docking Task: On a complex task involving fitting molecules into protein pockets (docking), SEGO found 10 hits in roughly half the number of trips required by existing approaches.

What SEGO Does Not Do

It's important to know what this method doesn't do, based on the paper's findings:

It doesn't work with random lists: If you just throw a random list of 500 molecules at the mapmaker (like a standard Bayesian Optimization approach), it fails to find hits because the list doesn't contain the right candidates. The mapmaker needs the scout to generate the candidates first.
It doesn't work with a "pure" compass: If you just let the generative scout spin out molecules without the mapmaker's guidance (a pure Reinforcement Learning approach), it fails to find hits under strict fuel limits. The scout gets lost without the map.
It doesn't guarantee infinite variety: The paper notes a side effect called "isomer collapse." Because the team is so good at finding the best treasure, they sometimes keep finding different versions of the same treasure (molecules with the same formula but different shapes). To fix this, the team had to add a special filter to force them to look for slightly different formulas, balancing the hunt for quality with the need for variety.

How Sure Are We?

The authors are very confident in these results because they were measured on established benchmarks.

Simulated Results: The performance numbers (like finding 10 hits in half the budget) come from running these methods on computer simulations of chemical tasks. The paper explicitly states these are results from the PMO benchmark and a multiparameter docking task.
Real-World Potential: The paper suggests that because SEGO is so efficient, it could eventually be used for direct experimental feedback (sending real ships to real islands) rather than just relying on cheap, low-quality computer guesses. However, the paper presents this as a future possibility enabled by the efficiency gains, not as a completed real-world clinical trial.

In short, SEGO is a clever loop where a mapmaker guides a scout to the right spot, and the scout's discoveries make the map even better. It turns a slow, expensive treasure hunt into a fast, precise mission, finding the best molecules with a fraction of the fuel.

Technical Summary: Sample Efficient Generative Optimization for Molecular Design

Problem Statement
Molecular optimization in drug discovery, materials design, and catalysis requires navigating vast chemical spaces under strict evaluation budgets. High-fidelity oracles (e.g., experimental synthesis and testing) are costly and time-consuming, making the "sample efficiency" of an optimization method—the number of evaluations required to find strong candidates—the primary determinant of its practical utility. Current methods face a trade-off: deep generative models (e.g., RL-based SMILES generators) often require hundreds to thousands of oracle calls to converge, while Bayesian Optimization (BO) offers high sample efficiency but is constrained by fixed libraries or sensitive to molecular representation choices. Existing hybrid approaches struggle to balance the flexibility of generative search with the data efficiency of BO, particularly in scenarios where the target distribution is unknown and direct experimental feedback is the only viable oracle.

Methodology: The SEGO Framework
The authors propose Sample Efficient Generative Optimization (SEGO), a tandem framework that integrates Reinforcement Learning (RL) and Bayesian Optimization (BO) to perform sample-efficient search in open-ended molecular space. SEGO operates through an iterative loop involving three core components:

Generative Engine (Saturn): An autoregressive SMILES generator based on the Mamba architecture, trained via RL. It utilizes "Augmented Memory," maintaining a replay buffer of top-rewarded SMILES strings that are randomized (via SMILES augmentation) to improve generalization and prevent mode collapse.
Bayesian Optimization Engine (GOLLuM): A surrogate model based on a Deep Kernel Gaussian Process (GP) that operates directly on SMILES embeddings derived from Large Language Models (LLMs). This allows for on-the-fly featurization without hand-crafted descriptors.
The SEGO Loop:
- Anchored Inner Loop: At each iteration, a "surrogate-guided" inner agent ( $\pi_{Inner}$ ) is instantiated from a "true-oracle" anchored agent ( $\pi_{Agent}$ ). The inner agent is trained against the GP posterior mean (the surrogate's hypothesis of where hits lie) to generate a targeted library of candidates in promising regions.
- Selection: The library is featurized using the deep-kernel GP, and an acquisition function (Expected Improvement) selects the most promising candidate.
- Feedback & Update: The selected candidate is evaluated by the true oracle. The result is used to:
  - Update the GP surrogate (refining the hypothesis).
  - Update the anchored agent ( $\pi_{Agent}$ ) via RL with the true reward.
  - Perform Augmented Optimization: The selected molecule is augmented into $N$ randomized SMILES variants, and the GP is trained on all representations simultaneously. This leverages the non-injective nature of SMILES to align the latent space, improving predictive accuracy, especially in low-data regimes.

Key Contributions

Surrogate-Guided Generation: Unlike standard BO which selects from fixed libraries, SEGO steers a generative model to propose candidates specifically in regions hypothesized by the surrogate to contain high rewards.
Anchoring Mechanism: The framework maintains an agent trained on true oracle feedback to ensure stability, while using a temporary inner agent to explore the surrogate's landscape. This prevents the agent from drifting into unproductive regions if the surrogate is initially uncertain.
SMILES-Augmented Surrogate Updates: By training the GP on multiple augmented representations of the same molecule, SEGO improves the alignment of the latent space and the accuracy of the surrogate model with fewer oracle calls.
Isomer Collapse Mitigation: The authors introduce a modified Expected Improvement acquisition function that inflates the incumbent threshold, encouraging the discovery of novel molecular formulas rather than just constitutional isomers of the current best hit.

Results

Ablation Studies: The anchored initialization strategy significantly outperforms re-initializing from a prior or maintaining a continuous agent across iterations, providing a necessary "warm start" in promising chemical regions. Augmented optimization (5x to 20x) was found to improve yield and reduce oracle burden without statistically significant loss in diversity.
Component Comparison: In a multi-property optimization task with a strict 100-call budget, the standalone generative engine (Saturn) and the standalone BO engine (GOLLuM) failed to find hits or performed poorly. SEGO successfully identified 63 unique hits, demonstrating the necessity of the hybrid approach.
PMO Benchmark: On the Practical Molecular Optimization (PMO) benchmark, SEGO achieved state-of-the-art performance using only 10% of the oracle budget (100 calls) consumed by other methods (which typically use 1,000 calls). It attained competitive top-1 and top-10 scores.
Multiparameter Docking: On a specific docking task, SEGO reached ten hits in roughly half the oracle calls required by existing state-of-the-art approaches.

Significance and Claims
The paper claims that SEGO moves molecular optimization closer to campaigns driven by direct experimental feedback. By drastically reducing the number of required oracle calls, the method makes it feasible to use high-fidelity simulations or real-world experiments as the optimization oracle, rather than relying on inexpensive, low-fidelity in silico proxies that are prone to "hacking." The authors position SEGO as a solution that overcomes the limitations of fixed libraries and the data inefficiency of pure generative methods, enabling the discovery of high-quality candidates in open-ended chemical spaces with minimal experimental cost.

The Scout Team Strategy

Why It's a Game-Changer

What SEGO Does Not Do

How Sure Are We?

Technical Summary: Sample Efficient Generative Optimization for Molecular Design

More like this